Replace builtin computer control with managed Cua MCP

AGENTS.md

@@ -0,0 +1,33 @@
+# Convera Agent Instructions
+
+## Real Application Verification
+
+For changes that affect application behavior, Electron main/preload/renderer
+integration, local AI providers, MCP servers or tools, settings, or
+user-visible flows, unit tests, lint, and typecheck are not sufficient by
+themselves.
+
+Before reporting the work complete:
+
+1. Read `packages/app/automation/README.md` completely.
+2. After a clean checkout or any app bundle/source change, run
+   `pnpm --filter @convera/app automation:prepare`.
+3. Start the automation MCP with `pnpm --filter @convera/app automation` and
+   drive the real Electron app with `convera_session`, `convera_observe`,
+   `convera_interact`, and `convera_wait`. Use `convera_execute` only when the
+   semantic actions cannot express the required step.
+4. Use an isolated `profile_id` and verify the production-facing behavior,
+   including visible success and the relevant failure or degraded state. Do
+   not substitute mocks or unit tests for this verification.
+5. Inspect the actual UI state and capture a screenshot or log artifact under
+   the ignored `packages/app/.automation/` directory when it provides useful
+   evidence.
+6. Close the automation session and clean up only the exact temporary
+   configuration and profile data created by the test. Never modify a real
+   user profile, session memo, or conversation database.
+7. Report the exercised scenario, provider/runtime, evidence, and any remaining
+   real-machine boundary. For provider work, do not run real Claude unless the
+   user explicitly asks; use real Codex when requested or available.
+
+If real application automation cannot run, do not silently call the feature
+complete. Report the blocker and the exact unverified boundary.

packages/app/package.json

@@ -95,7 +95,7 @@
     "@hurdlegroup/robotjs": "^0.12.3",
     "@icons-pack/react-simple-icons": "^12.2.0",
     "@leeoniya/ufuzzy": "^1.0.18",
-    "@modelcontextprotocol/sdk": "1.12.3",
+    "@modelcontextprotocol/sdk": "1.13.0",
     "@radix-ui/react-accordion": "^1.2.4",
     "@radix-ui/react-alert-dialog": "^1.1.7",
     "@radix-ui/react-aspect-ratio": "^1.1.3",

packages/app/src/electron/ai/__tests__/claude-code-mcp.test.ts

@@ -0,0 +1,84 @@
+import type { LocalAIChatRequest } from "@/shared/types/local-ai";
+import type { LanguageModel } from "ai";
+import { describe, expect, it, vi } from "vitest";
+import { LOCAL_AI_PROVIDER_DESCRIPTORS } from "../provider-descriptors";
+import { ClaudeCodeAdapter } from "../providers/claude-code";
+import type { LocalAiProviderStatus } from "../types";
+
+const mocks = vi.hoisted(() => {
+  const model = {} as LanguageModel;
+  const provider = vi.fn(() => model);
+
+  return {
+    model,
+    provider,
+    createClaudeCode: vi.fn(() => provider),
+    createSdkMcpServer: vi.fn(),
+    tool: vi.fn(),
+  };
+});
+
+vi.mock("ai-sdk-provider-claude-code", () => ({
+  createClaudeCode: mocks.createClaudeCode,
+  createSdkMcpServer: mocks.createSdkMcpServer,
+  tool: mocks.tool,
+}));
+
+describe("ClaudeCodeAdapter MCP transport", () => {
+  it("passes connected managed MCP servers to Claude without converting their tools", async () => {
+    const adapter = new ClaudeCodeAdapter();
+    const request: LocalAIChatRequest = {
+      requestId: "native-mcp",
+      conversationId: "conversation",
+      turnId: "turn",
+      providerId: "claude-code",
+      modelId: "claude-test",
+      operation: {
+        kind: "append",
+        message: { role: "user", content: "use cua" },
+      },
+      options: { cwd: "/tmp/convera-test" },
+    };
+    const status: LocalAiProviderStatus = {
+      ...LOCAL_AI_PROVIDER_DESCRIPTORS["claude-code"],
+      available: true,
+      authenticated: true,
+      executablePath: "/test/claude",
+      defaultModel: "claude-test",
+      models: ["claude-test"],
+      checkedAt: new Date(0).toISOString(),
+    };
+
+    const run = await adapter.prepareRun(request, status, {
+      tools: [],
+      nativeMcpServers: {
+        cua: {
+          transport: "stdio",
+          command: "cua-driver",
+          args: ["mcp"],
+          toolNames: ["screenshot"],
+        },
+      },
+      requestInteraction: vi.fn(async () => ({ approved: false })),
+    });
+
+    expect(run.model).toBe(mocks.model);
+    expect(mocks.provider).toHaveBeenCalledWith(
+      "claude-test",
+      expect.objectContaining({
+        pathToClaudeCodeExecutable: "/test/claude",
+        cwd: "/tmp/convera-test",
+        mcpServers: {
+          cua: {
+            type: "stdio",
+            command: "cua-driver",
+            args: ["mcp"],
+          },
+        },
+        allowedTools: ["mcp__cua__screenshot"],
+      }),
+    );
+    expect(mocks.createSdkMcpServer).not.toHaveBeenCalled();
+    expect(mocks.tool).not.toHaveBeenCalled();
+  });
+});